itically. | accuracy considerably. In LAND SAT data more subtle color changes due to atmospheric 
“ areas conditions or change in species composition can also result in serious misclassifications, 
of the whereas an interpretation would still furnish good results. 
Fis fast .. In all cases it is important to know how far the signatures which describe the classes 
in a data are valid and which subset of possible features is less susceptible to systematic or random 
changes. Unfortunately the spectral signatures mainly used in classification are very 
sensitive to changes in the data collection conditions. Therefore it is notonly important 
S to develop efficient sampling procedures to fully describe the variation within a class, 
4 5 ; . but also to use additional techniques to verify that the statistical descriptions are really 
3 > e valid for the whole classified area. With the maximum likelihood algorithm a statistical 
T i Bin" analysis of the achieved propabilities would yield first clues, if the used values are not 
tene truly representative for the classes. 
Furthermore unsupervised techniques are very useful to gain more understanding of the 
‘and data variation and can very effectively improve the definition of training areas or 
sions or statistical values reducing considerably this bottleneck in supervised classification. 
. model 
CONCLUSIONS 
his inter- Tia 
o the Compared with photointerpretation the unsupervised classification can be regarded as 
e bes! the determination of the units which can be separated. The selection of training areas 
valuation: and statistical calculation correspond to the definition of an interpretation key, 
normally accompanied by field checks. The supervised classification can be compared 
with the actual mapping of the photointerpretation units. During all steps the photo- 
s common interpreter will incorporate other information, e.g. maps, and his experience. In 
of classification only limited additional information is used, e.g. a digital terrain model. 
pie So far mainly spectral features are used for classification. Texture features were applied 
s hag to a limited number of problems. The additional use of size, shape or pattern would 
ái expand the application of classification techniques to other problems and data. It could 
in different also increase the accuracy, since some of these features are less dependant on the data 
gn collection conditions and can more easily be extended to other areas. But also with 
m or these techniques the general problem to determine the appropriate values describing 
inear, 
the classes remains the same. 
To incorporate the deduction process into a computer assisted evaluation procedure 
inner data . 
’ seems still to be a long distance away. 
assification 
117 
ET oc CC RE N i VE e SE a EDO 
ES eem ues - 
A